Invisible Neighbors: How AI-Mediated Communication Can Steer Collective Opinion

*A structured English summary of arXiv:2605.16245 (Tsirtsis, Rawal, Russell, Mittelstadt & Wachter), as discussed in the original Chinese forum post.*

Key points

• LLMs quietly rewrite stances. Four open-source LLM families — Llama-3.1-8B, Ministral-3-8B, Gemma-3-12B, Qwen3-8B — were tested on drafting and improving social media posts across 13 topics (UKP: abortion, cloning, death penalty, gun control, marijuana legalization, minimum wage, nuclear power, school uniforms; SemEval: abortion, atheism, climate change, feminism, Hillary, Trump). Despite system prompts instructing them to preserve the author's viewpoint and tone, all four models introduced directional bias on all 13 topics, measured on a continuous 0 (oppose) to 1 (support) scale.
• Bias correlates with the model's own views — mostly. The direction a model pulls posts during polishing generally matches the stance it expresses when asked directly. Exceptions exist (e.g., on atheism, a model expressed positive attitudes yet pulled edits toward opposition), showing that simple direct probing fails to capture biases in real interactions.
• Networks amplify one-off bias into collective drift. The authors insert an AI transformation function into the classic Friedkin-Johnsen opinion dynamics model:

x_i(t+1) = λ_i · x_i(0) + (1-λ_i) · Σ_j W_ij · f(x_j(t))

For a linear transformation f_lin(x) = mx + (1-m)ν, they derive closed-form propositions: the system still converges to a unique equilibrium exponentially fast; the equilibrium shift depends on network structure W, not just AI parameters; and under homogeneous stubbornness with doubly stochastic W, the mean opinion shift equals a multiplicative amplification of the one-off AI bias — strictly greater than 1 when m(1-λ) > λ.

• Simulations on real networks are striking. Using SNAP Twitter network data (~80,000 nodes, 1.7M edges) and Gemma-3's non-linear transformations (estimated via Nadaraya-Watson kernel regression):
• On abortion, the average opinion — originally opposed — flips to supportive when 60% of users use AI editing.
• The ratio of long-run equilibrium shift to one-off AI bias reaches up to 9.2x.
• Amplification is strongest when the population is nearly split and the AI favors the minority side, potentially converting a minority into a majority.
• Case study: X's Grok "Explain this post." Reproducing X's publicly released prompt (https://github.com/xaiorg/grokprompts/blob/main/grok_analyze_button.j2), the authors ran 39 pro-choice and 39 pro-life posts through Grok five times each (1,170 claims), classified by GPT-5.4. Results: β_sup = 0.24 (Grok echoes pro-life posts much more than pro-choice posts) vs. β_opp = 0.04, both statistically significant. An ablation over the prompt's four guidelines shows the third one — "provide truthful, insightful, and factual perspectives, challenging mainstream narratives" — is solely responsible: removing it eliminates both biases. The authors call this "Bias by Design": a platform design choice, not an emergent model quirk.
• Regulatory gap. Under the EU AI Act, Article 53 imposes no obligation to test or disclose stance bias in text polishing; Article 55's systemic-risk threshold (red-teaming, incident reporting) likely isn't met; Annex III(8)(b) targets systems aimed at influencing voting, which "improve my post" features explicitly are not; Article 50(4)'s transparency rules mostly exclude AI-generated text. Under the DSA Article 34, proving such bias as a systemic risk requires a causal chain not yet established in enforcement practice — and enforcement is already strained (platforms withdrawing from independent data-access commitments; X fined €120M). The paper concludes existing frameworks leave AI-mediated opinion steering essentially unregulated.

Intuition: ink on filter paper

The post offers a Feynman-style metaphor: a drop of ink spreads slowly on flat paper, but if the paper is a network of threads each pulling the ink slightly in one direction, the whole sheet eventually takes one color. AI acts as an invisible neighbor sitting between every pair of communicators, micro-adjusting each outgoing and incoming message — each adjustment imperceptible, even welcome — but collectively steering society's consensus toward the model's "neutral point." The danger is not persuasion by argument but invisible steering of expression itself.

References

• Tsirtsis, S., Rawal, K., Russell, C., Mittelstadt, B., & Wachter, S. (2026). AI-Mediated Communication Can Steer Collective Opinion. arXiv:2605.16245 [cs.CY].
• DeGroot, M. H. (1974). Reaching a consensus. *JASA*, 69(345), 118-121.
• Friedkin, N. E., & Johnsen, E. C. (1999). Social influence networks and opinion change. *Advances in Group Processes*, 16(1), 1-29.
• Stabczyk, A., et al. (2026). Large language models reflect the ideology of their creators. *npj Artificial Intelligence*, 2(1), 7.